The hardest problem in deploying autonomous drones has never been making one fly well. It is making it fly well on a drone it was not built for. A paper published in Science Robotics this month has a solution that fits in a file smaller than most email attachments.
Every drone AI today is a specialist. Train it on one machine and it flies that machine well. Give it a drone with slightly different motors, a heavier frame, or a different propeller count, and it fails. Engineers have to retrain or retune for each new platform. In an industry spanning 30-gram hobby craft to multi-kilogram delivery vehicles, that is a serious constraint on how widely autonomous drone technology can be deployed.
RAPTOR, developed by researchers at UC Berkeley and the Technology Innovation Institute in Abu Dhabi, is a single neural network controller that flies any quadrotor it has never previously encountered. It adapts in milliseconds, with no retraining and no prior knowledge of the specific machine it has been loaded onto. Published in Science Robotics on 13 May, it was tested on 10 real drones the policy had never seen, from a 32-gram micro-drone to a 2.4-kilogram platform.
How it works
The researchers trained 1,000 separate AI controllers, one for each of 1,000 simulated drones with randomised physical properties. They then compressed all 1,000 into a single controller through a process called meta-imitation learning. The result learned not just how to fly, but how to figure out what kind of drone it was on, by watching how the drone responded to its first few commands.
This is called in-context learning. Rather than being told a drone’s specifications, RAPTOR infers them in real time from behaviour. If the drone responds sluggishly to a throttle increase, the controller concludes it is heavy. If it overshoots turns, it compensates. The adaptation takes milliseconds.
The network that does all of this has 2,084 parameters. The AI model behind ChatGPT has hundreds of billions.
What they tested
The 10 test drones varied in motor type (brushed versus brushless), frame construction (rigid versus flexible), propeller count (two, three, or four blades), and flight control software, covering PX4, Betaflight, Crazyflie, and M5StampFly. Tests ran indoors and outdoors, in wind, with deliberate mid-flight disturbances, and with mismatched propellers.
One unchanged policy maintained stable flight and accurate trajectory tracking across every combination.
Why this matters
Serious autonomous drone applications, whether delivery, infrastructure inspection, or search-and-rescue, require controllers developed and validated for each vehicle type. Add a new aircraft to a fleet, and the AI work starts again. RAPTOR removes that bottleneck. A policy that adapts on contact with the hardware means new drone models can join autonomous fleets without rebuilding the AI from scratch.
The broader significance is what the approach suggests about robotics AI generally. The paper is part of a growing effort to create “foundation policies” for physical robots: train on enough diversity, compress the result, and build a system that adapts from context. RAPTOR is one of the first demonstrations that this works across genuinely different physical hardware in the real world, not just in simulation.
What comes next
The 10 test drones are a subset of quadrotor designs in commercial use, and testing covered trajectory tracking rather than complex mission scenarios. The adaptation also assumes consistent drone behaviour in the first seconds of flight, which may not hold if hardware is damaged or degraded.
The policy weights and codebase are open source and built to drop into flight controller software already standard across the industry. That lowers the barrier from research result to practical trial considerably.
An AI drone controller with 2,084 parameters that fits any aircraft it is loaded onto is not a refinement of existing technology. It is a different assumption about how drone intelligence should be built.
Sources
Eschmann, J., Albani, D., & Loianno, G. (2026). RAPTOR: A Foundation Policy for Quadrotor Control. Science Robotics, 11, eaec1481. Published 13 May 2026. https://www.science.org/doi/10.1126/scirobotics.aec1481
Companion Materials
Project website: https://raptor.rl.tools Video: https://youtu.be/hVzdWRFTX3k (linked in the paper; posted October 2025, corresponds to the arXiv preprint version) Code repository (Zenodo): https://doi.org/10.5281/zenodo.17096679
More Like This
Ray Jackson holds a BSc in Electrical Engineering from the University of Manitoba and a PhD in Physics from Carleton University. His reporting interests include Current and Future Technologies, Engineering and Artificial Intelligence.